Architecture

Understanding TDP and GPU Power Consumption

TDP (Thermal Design Power) measures the maximum power a GPU draws under load, expressed in watts. GPU power consumption directly impacts data centre operating costs — a 700W H100 running 24/7 costs approximately $50/month in electricity alone. According to Signwl data, cloud GPU TDP ranges from 70W (T4) to 1,400W (GB300), with power efficiency (TFLOPS per watt) varying dramatically across generations.

What is TDP?

TDP represents the maximum sustained power draw a GPU is designed for. A GPU with a 700W TDP (like the H100) requires cooling infrastructure capable of dissipating 700W of heat.

Actual power consumption varies with workload — a GPU running inference may draw 40-60% of its TDP, while intensive training can push close to the rated TDP. Power draw also depends on GPU utilisation, clock speeds, and the specific operations being performed.

Power Efficiency Across Generations

Each GPU generation has improved power efficiency (TFLOPS per watt), but absolute power consumption has also increased as performance grows faster than efficiency.

Signwl tracks TFLOPS per watt across all GPU types. The T4 at 0.93 TFLOPS/W and the L4 at 1.68 TFLOPS/W are among the most efficient. The H100 delivers 1.41 TFLOPS/W, while the GB300 achieves 1.61 TFLOPS/W despite its massive 1,400W TDP.

For cloud users, power consumption is already factored into hourly pricing — providers include electricity costs in their rates. However, understanding TDP helps when planning on-premises deployments or evaluating total cost of ownership.

Power and Data Centre Design

GPU power consumption is a major factor in data centre design and location. Modern GPU clusters require substantial electrical infrastructure and cooling capacity.

This is driving interest in regions with low electricity costs and cool climates. Nordic countries, parts of Canada, and certain US states attract GPU deployments partly due to favourable power economics.

Signwl's regional pricing data reflects these power cost differences — regions with expensive electricity tend to have higher GPU hourly rates.

Frequently Asked Questions

What is TDP?

TDP (Thermal Design Power) is the maximum power a GPU draws under sustained load, measured in watts. The H100 has a 700W TDP, meaning it requires cooling for up to 700W of heat dissipation.

Which GPU is the most power efficient?

For AI workloads, the NVIDIA L4 offers the best TFLOPS per watt at 1.68 TFLOPS/W with only 72W TDP. The T4 is also highly efficient at 0.93 TFLOPS/W with 70W TDP. Among training GPUs, the GB300 achieves 1.61 TFLOPS/W.

Does power consumption affect cloud GPU pricing?

Yes — cloud providers include electricity costs in hourly rates. Regions with higher electricity costs tend to have higher GPU pricing. This is partly why North American pricing is generally lower than European or Asian pricing.